Timezone: »

 
On the Global Convergence of the Regularized Generalized Gauss-Newton Algorithm
Vincent Roulet · Maryam Fazel · Siddhartha Srinivasa · Zaid Harchaoui

@
in Order up! The Benefits of Higher-Order Optimization in Machine Learning »

We detail the global convergence rates of a regularized generalized Gauss-Newton algorithm applied to compositional problems with surjective inner Jacobian mappings. Our analysis uncovers several convergence phases for the algorithm and identifies the key condition numbers governing the complexity of the algorithm. We present an implementation with a line-search adaptive to the constants of the problem.

[ Poster] 

Author Information

Vincent Roulet (UW)
Maryam Fazel (University of Washington)
Siddhartha Srinivasa (University of Washington)
Zaid Harchaoui (University of Washington)

More from the Same Authors